MiniMax Music 3 GGUF

MiniMax Music 3 generates complete songs up to 5 minutes long at 32 kHz stereo. It uses a 2.4B flow-matching diffusion transformer plus a text and language model encoder. This repository ships GGUF quants of both components.

Hugging Face repositoryMiniMax-Music3-GGUF on Hugging Face

Example output

Available Quantizations

Diffusion transformer (DiT)

Quant	Size	File
BF16	4.64 GB	minimax_music3_dit_BF16.gguf
Q4_0	1.39 GB	minimax_music3_dit_Q4_0.gguf
Q8_0	2.52 GB	minimax_music3_dit_Q8_0.gguf
Q8_CR	2.38 GB	minimax_music3_dit_Q8_CR.gguf

Pruned text encoder

Quant	Size	File
Q4_0	5.38 GB	minimax_music3_text_encoder_pruned_Q4_0.gguf
Q8_0	8.93 GB	minimax_music3_text_encoder_pruned_Q8_0.gguf
Q8_CR	8.49 GB	minimax_music3_text_encoder_pruned_Q8_CR.gguf

Model Input and Output

Inputs

Field	Value
Lyrics	Text with uppercase section tags
Music description	Genre, tempo, key, vocals, arrangement
Token limit	5,000 tokens

Outputs

Field	Value
Duration	Up to 5 minutes
Audio	32 kHz stereo
Streaming	No

ComfyUI Setup

Load the model with the GGUF loader node from comfyui-gguf-reboot. The loader node gives the best performance. Use Q8_CR for the DiT when audio quality matters.

Dependencies

Place these files in the ComfyUI model folders before running a workflow.

Component	File	Folder
Diffusion transformer	minimax_music3_dit_*.gguf	models/diffusion_models/ or models/unet/
Text encoder	minimax_music3_text_encoder_pruned_*.gguf	models/text_encoders/ or models/clip/

Workflow downloads

Import this JSON file into ComfyUI.

JSON fileMusic generation ComfyUI workflow